You Order Newsprint in July for a Football Tab You Cannot Size Yet. 2026 Forecasting Models Set the Draw Rack by Rack.

Nine weeks. That is roughly how far ahead a community paper commits to newsprint for the fall — you are placing the order in early July for the football preview tab that runs in the third week of August, the fair edition, and the first three home-game Fridays. In July you do not know whether the varsity team is going to be 8‑1 or 1‑8, whether the district will finally break ground on the elementary addition, or whether the mill is going to announce layoffs. You order anyway. Then you set the draw — the number of copies you actually print — edition by edition, and you find out on Tuesday whether you guessed right by counting what came back from the racks.

Most publishers guess high on the tabs and low in January, because a sellout feels worse than a pallet of returns. That instinct is expensive in both directions, and it has gone unexamined for forty years because nothing could make a decent forecast out of the history a small publisher actually has: three years of spotty draw-and-returns by outlet, an edition calendar that shifts with the school year, and a dozen one-off weeks that blow up every average.

Why is the demand in this business so lumpy?

Because almost none of it is driven by the calendar alone. A weekly's single-copy demand is driven by events: the football tab in mid-August, the county fair edition, the graduation and senior salute section in May, the progress edition in January or February, the delinquent tax list whose publication date your state statute sets, sample ballots and election notices in even-numbered years, and the two or three weeks a year when a well-known local figure dies or the school board fires a superintendent and every rack in town clears by noon.

Advertising demand is just as lumpy and runs on a different clock: holiday preprints and inserts piling up through November, back-to-school in August, home and garden in April, and the annual pattern where three of your top ten accounts spend nothing at all in the first six weeks of the year. Niche and trade titles have their own version — the show issue with bonus distribution, the buyer's guide, the ad close date that sits three weeks ahead of the materials close.

Demand forecasting is now one of the most widely adopted uses of AI in industry — around 48 percent adoption in manufacturing — and pricing optimization sits around 72 percent in retail and e-commerce, precisely because 2026 models handle messy, sparse, seasonal history that older statistical methods choked on. Put plainly: a forecast that once needed five clean years of weekly data can now be built from what a 6,400-circulation weekly already has sitting in Newzware or Naviga Circulation, gaps and all.

flowchart TD
  A["3 years of draw and returns, rack by rack"] --> D["Forecast next 9 weeks"]
  B["Last 6 football tabs and fair editions"] --> D
  C["School calendar, fair dates, election dates"] --> D
  D --> E{"Rack sell-through forecast above 85%?"}
  E -->|Yes| F["Raise draw at those 6 stores"]
  E -->|No| G["Cut draw at the 11 chronic-return stores"]
  F --> H["Newsprint order to the mill"]
  G --> H

What does guessing the draw actually cost?

Two ways, and only one of them shows up on a statement. Over-printing costs newsprint, ink, press time, the haul, and the labor to count and pulp returns — a real per-copy number your press foreman can give you in about four minutes. Under-printing costs the single-copy sale you did not make, and something worse: the reader who drove to the IGA at 7 a.m., found an empty rack, and did not drive to the next one. Do that three times to the same person and you have donated a subscriber to nobody.

There is a third cost publishers rarely price. Returns concentrate at the same eleven outlets every week, usually the convenience stores nobody has re-audited since the last circulation manager left. You have been printing, delivering and retrieving copies to those racks for years because nobody had time to pull the numbers store by store.

Setting price when your history is thin

The same models are being pointed at the other half of the problem: what you charge. Most small publishers run one January increase across the whole subscriber file — everyone up $1.50 a month, brace for the stop calls in the first two weeks of February, and hope the net is positive. Nobody knows the net, because nobody separates the people who would have stopped anyway from the people the increase pushed out.

What a 2026 model does with your renewal history is separate the file into groups by behavior you already recorded: how long they have taken the paper, whether they are on EZ Pay auto-renew or pay by check, in-county mail versus motor route versus digital-only, how many times they have called about a missed delivery, and whether they took the paper the last time you raised the price. Then it tells you which groups tolerate the increase and which ones you should leave alone this year. The same logic works on the rate card — the preprint insert rate per thousand, the ROP display rate for a 2-column by 5-inch, the obituary rate with and without a photo, and where discounting is buying nothing.

A worked example on one fall's newsprint

Assume a weekly with a 6,400 draw, single copy at 2,900 of that across 42 outlets, mailed subscribers at 3,500. Assume returns currently run 18 percent of single copy, and a fully loaded cost of 41 cents a copy for newsprint, ink, press time and hauling. Assume the forecast lets you hold sell-through steady while cutting the chronic-return outlets, bringing returns to 11 percent.

Two honest notes on that table. The 41 cents is the number you should replace with your own; press economics at a shop that prints its own paper look nothing like a shop that buys press time from the daily 60 miles away. And the sellout recovery line is the softer of the two — you are estimating copies you did not sell, which is always partly a story you tell yourself. If you only believe the first number, $4,162 a year still pays for the software several times over.

What the forecast cannot see coming

It cannot see the superintendent resigning on a Thursday night. It cannot see the fatal accident on the county highway, the tornado, the plant closure, or the arrest of somebody's cousin who happens to be a village trustee. Those are the weeks a weekly earns its place in town, and they will always be a phone call from your editor to your press foreman at 9:40 p.m. saying add 400.

It also cannot decide whether you should be printing at all in a given market, or whether the right move is to convert five chronic-return racks into a digital-only push. That is a strategy call about what your paper is for, and it belongs to you. And be careful with the pricing side: a model optimizing purely for revenue per subscriber will happily recommend an increase that empties your in-county mailed list, and your in-county paid subscribers are part of what qualifies you to carry public notices in most states. Ask the forecast for options; keep the qualification test in your own head.

Start with 42 racks and one spreadsheet

Pull your last 156 weeks of draw and returns by outlet out of your circulation system, however ugly the export is, and hand the whole file to Claude Cowork or ChatGPT Work with a plain instruction: rank the outlets by returns as a share of draw, show me which weeks broke the pattern, and tell me what you would set next week's draw to at each one. You are not committing to anything. You are finding out whether the machine sees the eleven stores you already suspect. If it does, you have a forecasting tool. If it names eleven stores you did not suspect, you have something better.

Frequently asked questions

Our circulation data is a mess and half of it is in a spreadsheet the last manager kept. Is that a dealbreaker?

No, and that is the actual news. The reason this became practical in 2026 is that the current models cope with sparse, gappy, inconsistent history — the exact kind small publishers keep. Missing weeks and renamed outlets are handled; you just want the model to tell you where it is uncertain instead of smoothing over it.

Will a forecast tell me whether to raise the newsstand price?

It will tell you what happened at your own outlets the last time you moved it, and what similar patterns in your file suggest. It will not tell you what the drug store in the next town over charges, or how the county feels about paying two dollars. Cover price is part arithmetic, part politics; use the model for the first half.

How far out is a forecast for a weekly actually useful?

Roughly the length of your newsprint lead time, which is why nine weeks is the number that matters. Beyond a quarter you are guessing at events. Inside two weeks you already know most of what you need from the ad book and the school calendar.

Does this touch advertising rates too, or just circulation?

Both, though the ad side takes more care. The clean win is spotting which accounts have quietly slid from 26 insertions a year to 14 and pricing a package to bring them back before they lapse entirely. The bad use is letting the model set spot rates for a market of 60 advertisers who all talk to each other at the Chamber lunch.

One note on the phone

Forecasting fixes the draw; it does not answer the calls a sellout week generates — readers asking which store still has copies, subscribers calling because they want the fair edition held, and advertisers calling to get into a section that just got hot. Those come in a burst on Wednesday morning and land on whoever is at the front counter. CallSphere builds AI voice and chat agents that answer the line and the website chat, capture the request and pass it to your circulation manager, so a big week does not cost you the calls you could not pick up. It will not set your draw — that stays between your numbers and your press foreman.